/
hparams.py
222 lines (192 loc) · 5.04 KB
/
hparams.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
# coding: utf-8
import tensorflow as tf
import numpy as np
from os.path import join, dirname
def hparams_debug_string(params):
values = params.values()
hp = [' %s: %s' % (name, values[name]) for name in sorted(values)]
return 'Hyperparameters:\n' + '\n'.join(hp)
# Hyper parameters for voice conversion
vc = tf.contrib.training.HParams(
# Acoustic features
order=59,
frame_period=5,
windows=[
(0, 0, np.array([1.0])),
(1, 1, np.array([-0.5, 0.0, 0.5])),
(1, 1, np.array([1.0, -2.0, 1.0])),
],
static_dim=59,
stream_sizes=None,
has_dynamic_features=None,
adversarial_streams=None,
# Generator
generator="In2OutHighwayNet",
generator_params={
"in_dim": None,
"out_dim": None,
"num_hidden": 3,
"hidden_dim": 512,
"static_dim": 59,
"dropout": 0.5,
},
optimizer_g="Adagrad",
optimizer_g_params={
"lr": 0.01,
"weight_decay": 0,
},
# Discriminator
discriminator="MLP",
discriminator_params={
"in_dim": 59,
"out_dim": 1,
"num_hidden": 2,
"hidden_dim": 256,
"dropout": 0.5,
"last_sigmoid": True,
},
optimizer_d="Adagrad",
optimizer_d_params={
"lr": 0.01,
"weight_decay": 0,
},
# This should be overrided
nepoch=200,
# LR schedule
lr_decay_schedule=False,
lr_decay_epoch=10,
# Datasets and data loader
batch_size=32,
num_workers=1,
pin_memory=True,
cache_size=1200,
)
# Hyper paramters for TTS duration model
tts_duration = tf.contrib.training.HParams(
# Linguistic features
use_phone_alignment=False,
subphone_features=None,
add_frame_features=False,
question_path=join(dirname(__file__), "nnmnkwii_gallery", "data",
"questions-radio_dnn_416.hed"),
# Duration features
windows=[
(0, 0, np.array([1.0])),
],
stream_sizes=[5],
has_dynamic_features=[False],
# Streams used for computing adversarial loss
adversarial_streams=[True],
mask_0th_mgc_for_adv_loss=False,
# Generator
generator="MLP",
generator_params={
"in_dim": None,
"out_dim": None,
"num_hidden": 3,
"hidden_dim": 512,
#"bidirectional": True,
"dropout": 0.5,
"last_sigmoid": False,
},
optimizer_g="Adagrad",
optimizer_g_params={
"lr": 0.01,
"weight_decay": 1e-7,
},
# Discriminator
discriminator="MLP",
discriminator_params={
"in_dim": None,
"out_dim": 1,
"num_hidden": 2,
"hidden_dim": 256,
"dropout": 0.5,
"last_sigmoid": True,
},
optimizer_d="Adam",
optimizer_d_params={
"lr": 0.01,
"weight_decay": 1e-7,
},
# This should be overrided
nepoch=200,
# LR schedule
lr_decay_schedule=False,
lr_decay_epoch=25,
# Datasets and data loader
batch_size=32,
num_workers=1,
pin_memory=True,
cache_size=1200,
)
# Hyper paramters for TTS acoustic model
tts_acoustic = tf.contrib.training.HParams(
# Linguistic
use_phone_alignment=False,
subphone_features="full",
add_frame_features=True,
question_path=join(dirname(__file__), "nnmnkwii_gallery", "data",
"questions-radio_dnn_416.hed"),
# Acoustic features
order=59,
frame_period=5,
windows=[
(0, 0, np.array([1.0])),
(1, 1, np.array([-0.5, 0.0, 0.5])),
(1, 1, np.array([1.0, -2.0, 1.0])),
],
# Stream info
stream_sizes=[180, 3, 1, 3],
has_dynamic_features=[True, True, False, True],
# Streams used for computing adversarial loss
# NOTE: you should probably change discriminator's `in_dim`
# if you change the adv_streams
adversarial_streams=[True, True, False, False],
# Don't switch this on unless you are sure what you are doing
# If True, you will need to adjast `in_dim` for discriminator.
# Rationale for this is that power coefficients are less meaningful
# to distinguish natrual/generated, especially for frame-level models.
mask_0th_mgc_for_adv_loss=True,
# Generator
generator="MLP",
generator_params={
"in_dim": None,
"out_dim": None,
"num_hidden": 3,
"hidden_dim": 512,
#"bidirectional": True,
"dropout": 0.5,
"last_sigmoid": False,
},
optimizer_g="Adagrad",
optimizer_g_params={
"lr": 0.01,
"weight_decay": 1e-7,
},
# Discriminator
discriminator="MLP",
discriminator_params={
"in_dim": 60,
"out_dim": 1,
"num_hidden": 2,
"hidden_dim": 256,
"dropout": 0.5,
"last_sigmoid": True,
},
optimizer_d="Adagrad",
optimizer_d_params={
"lr": 0.01,
"weight_decay": 1e-7,
},
# This should be overrided
nepoch=200,
# LR schedule
lr_decay_schedule=False,
lr_decay_epoch=25,
# Datasets and data loader
batch_size=26,
num_workers=1,
pin_memory=True,
cache_size=1200,
)